---
title: "The Danger of Overthinking — paper by Cuadron et al. (2025)"
description: "Large Reasoning Models (LRMs) such as o1, QwQ and DeepSeek-R1 perform well on static… 34 references, cited by 39 papers here."
authors: ["Alejandro Cuadron", "Dacheng Li", "Wenjie Ma", "Xingyao Wang", "Yichuan Wang", "Siyuan Zhuang", "Shu Liu", "Luis Gaspar Schroeder", "Tian Xia", "Huanzhi Mao", "Nicholas Thumiger", "Aditya Desai", "Ion Stoica", "Ana Klimovic", "Graham Neubig", "Joseph E. Gonzalez"]
year: 2025
venue: "arXiv:2502.08235"
source: "https://arxiv.org/abs/2502.08235"
doi: "10.48550/arxiv.2502.08235"
arxiv: "2502.08235"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/cuadron-2025-danger-overthinking-examining-reasoning-action/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/cuadron-2025-danger-overthinking-examining-reasoning-action/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/3/) / The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks

# The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks

**Authors:** [Alejandro Cuadron](https://md.research.elamur.ai/free/en/authors/alejandro-cuadron/), [Dacheng Li](https://md.research.elamur.ai/free/en/authors/dacheng-li/), [Wenjie Ma](https://md.research.elamur.ai/free/en/authors/wenjie-ma/), [Xingyao Wang](https://md.research.elamur.ai/free/en/authors/xingyao-wang/), Yichuan Wang, [Siyuan Zhuang](https://md.research.elamur.ai/free/en/authors/siyuan-zhuang/), [Shu Liu](https://md.research.elamur.ai/free/en/authors/shu-liu/), Luis Gaspar Schroeder, [Tian Xia](https://md.research.elamur.ai/free/en/authors/tian-xia/), [Huanzhi Mao](https://md.research.elamur.ai/free/en/authors/huanzhi-mao/), Nicholas Thumiger, Aditya Desai, [Ion Stoica](https://md.research.elamur.ai/free/en/authors/ion-gabriel-stoica/), Ana Klimovic, [Graham Neubig](https://md.research.elamur.ai/free/en/authors/graham-neubig/) and [Joseph E. Gonzalez](https://md.research.elamur.ai/free/en/authors/joseph-e-gonzalez/) · **Year:** 2025 · **Venue:** arXiv:2502.08235 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2502.08235) · [PDF](https://arxiv.org/pdf/2502.08235) · [DOI 10.48550/arxiv.2502.08235](https://doi.org/10.48550/arxiv.2502.08235) · [arXiv 2502.08235](https://arxiv.org/abs/2502.08235) · [OpenAlex](https://openalex.org/W4407570327)

**Citations:** 2 citations (OpenAlex, 2026-10-14) · [39](https://md.research.elamur.ai/free/en/papers/cuadron-2025-danger-overthinking-examining-reasoning-action/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
Large Reasoning Models (LRMs) such as o1, QwQ and DeepSeek-R1 perform well on static benchmarks. In agentic settings, however, they must keep choosing between acting in the environment and continuing to reason internally. The authors call this the Reasoning-Action Dilemma. They study "overthinking," where a model relies on long internal simulations instead of collecting real feedback.

Using SWE-bench Verified with the CodeAct agent inside OpenHands, they built an LLM-as-a-judge scoring framework and validated it against human expert ratings. It flags three distinct kinds of failure: Rogue Actions, Premature Disengagement, and Analysis Paralysis. They applied it to roughly 4,000 trajectories from reasoning and non-reasoning models. Higher overthinking scores went with lower issue-resolution rates, and reasoning models overthought considerably more.

A simple selection trick recovers most of the accuracy of expensive high-effort runs at much lower cost: sample several low-effort runs and keep the one with the lowest overthinking score. The work matters for anyone deploying reasoning models as agents, where both compute budgets and real interaction with tools matter.

## Key points
- **Benchmark and dataset.** The benchmark is SWE-bench Verified with OpenHands/CodeAct scaffolding. The paper gives the trajectory count as 4,018 in the abstract but 3,908 in the conclusion. The framework and dataset are open-sourced at github.com/AlexCuadron/Overthinking.
- **Reasoning vs. non-reasoning models.** Reasoning models score about three times higher on overthinking than non-reasoning models. Both groups show a significant negative regression between overthinking and resolution rate (R² = 0.892 and 0.839), with a steeper slope for non-reasoning models.
- **Cost vs. accuracy for o1.** With high reasoning effort, o1 resolves 29.1% of issues at $1,400. With low effort, it resolves 21.0% at $400.
- **Lowest Overthinking@k.** With k=2 low-effort samples, this method reaches 27.3% at $800, a 43% cost cut. With k=3, it reaches 30.3%, beating the high-effort baseline. The headline gain is stated as "almost 30%" in the abstract but 25% in the conclusion.
- **Proposed mitigations.** The authors suggest native function calling (o1 with FC performs notably better) and selective reinforcement learning. They also note that DeepSeek-R1-671B overthinks no more than DeepSeek-V3, which they attribute to limited software-engineering RL and its large scale.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) — University, Pittsburgh, United States · 633 works · 195,218 citations
- [University of California, Berkeley](https://md.research.elamur.ai/free/en/universities/university-of-california-berkeley/) — University, Berkeley, United States · 243 works · 251,774 citations
- [University of Illinois Urbana-Champaign](https://md.research.elamur.ai/free/en/universities/university-of-illinois-urbana-champaign/) — University, Urbana, United States · 228 works · 74,591 citations
- [ETH Zurich](https://md.research.elamur.ai/free/en/universities/eth-zurich/) — University, Zurich, Switzerland · 97 works · 9,612 citations

📚 Shelves (3): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 19,503 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 18,805 works · [🗄️ Information Systems](https://md.research.elamur.ai/free/en/shelves/computer-science/information-systems/) · 1,888 works

🧭 Topics (1): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,948 works

## Cite

Cite as: Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, Nicholas Thumiger, Aditya Desai, Ion Stoica, Ana Klimovic, Graham Neubig and Joseph E. Gonzalez. 2025. The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks. arXiv:2502.08235.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{cuadron2025danger,
  title = {{The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks}},
  author = {Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, Nicholas Thumiger, Aditya Desai, Ion Stoica, Ana Klimovic, Graham Neubig and Joseph E. Gonzalez},
  year = {2025},
  doi = {10.48550/arxiv.2502.08235},
  eprint = {2502.08235},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2502.08235}
}
```

- [← previous DeepSeek: Paradigm Shifts and Technical Evolution in Large AI Models](https://md.research.elamur.ai/free/en/papers/xiong-2025-deepseek-paradigm-shifts-technical/)
- [next → A Systematic Survey of Automatic Prompt Optimization Techniques](https://md.research.elamur.ai/free/en/papers/ramnath-2025-systematic-survey-automatic-prompt-optimization/)
- [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/3/)

## References

34 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Liu, T., Chang, B., Sun, X., Li, L., and Sui, Z. [A survey on in-context learning](https://md.research.elamur.ai/free/en/papers/dong-2023-survey-context-learning/), 2024. URL https://arxiv.org/abs/2301.00234. [↗](https://arxiv.org/abs/2301.00234)
2. Anthropic. Claude 3.5: A sonnet of progress in ai. [https://www.anthropic.com/news/](https://www.anthropic.com/news/) claude-3-5-sonnet, 2024. Accessed: 2024-11-21. [↗](https://www.anthropic.com/news/)
3. Guan, X., Zhang, L. L., Liu, Y., Shang, N., Sun, Y., Zhu, Y., Yang, F., and Yang, M. [rstar-math: Small llms can master math reasoning with self-evolved deep thinking](https://md.research.elamur.ai/free/en/papers/guan-2025-rstar-math-small-llms-can/), 2025. URL https://arxiv.org/abs/2501.04519. [↗](https://arxiv.org/abs/2501.04519)
4. AoPS. 2024 aime i. [https://artofproblemsolving.com/wiki/index](https://artofproblemsolving.com/wiki/index). php/2024_AIME_I, 2024. Accessed: 2025-01-22. [↗](https://artofproblemsolving.com/wiki/index)
5. ARC. Openai o3 publication breakthrough, 2024. URL [https://arcprize.org/blog/](https://arcprize.org/blog/) oai-o3-pub-breakthrough. Accessed: 2024-11-03. [↗](https://arcprize.org/blog/)
6. Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. [Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning](https://md.research.elamur.ai/free/en/papers/deepseek-ai-2025-deepseek-r1-incentivizing-reasoning-capability/). arXiv preprint arXiv:2501.12948, 2025.
7. Besta, M., Barth, J., Schreiber, E., Kubicek, A., Catarino, A., Gerstenberger, R., Nyczyk, P., Iff, P., Li, Y., Houliston, S., Sternal, T., Copik, M., Kwaśniewski, G., Müller, J., Flis, l., Eberhard, H., Niewiadomski, H., and Hoefler, T. Reasoning Language Models: A Blueprint, January 2025. [find](https://scholar.google.com/scholar?q=Besta%2C+M%2E%2C+Barth%2C+J%2E%2C+Schreiber%2C+E%2E%2C+Kubicek%2C+A%2E%2C+Catarino%2C+A%2E%2C+Gerstenberger%2C+R%2E%2C+Nyczyk%2C+P%2E%2C+Iff%2C+P%2E%2C+Li%2C+Y%2E%2C+Houliston%2C+S%2E%2C+Sternal%2C+T%2E%2C+Copik%2C+M%2E%2C+Kwas%CC%81niewski%2C+G%2E%2C+Mu%CC%88ller%2C+J%2E%2C+Flis%2C+l%2E%2C+Eberhard)
8. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. [Measuring massive multitask language understanding](https://md.research.elamur.ai/free/en/papers/hendrycks-2020-measuring-massive-multitask-language-understanding/), 2021. URL https://arxiv.org/abs/2009.03300. [↗](https://arxiv.org/abs/2009.03300)
9. Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, Blog, A. D. Reinventing the amazon q de-O., and Narasimhan, K. Swe-bench: Can language veloper agent for software development. models resolve real-world github issues?, 2024. URL — [SWE-bench: Can Language Models Resolve Real-World GitHub Issues?](https://md.research.elamur.ai/free/en/papers/jimenez-2024-swe-bench-can-language-models/) [https://aws.amazon.com/blogs/devops/](https://aws.amazon.com/blogs/devops/) https://arxiv.org/abs… [↗](https://aws.amazon.com/blogs/devops/)
10. A., Naumann, T., Globerson, A., Saenko, K., Chen, X., Xu, J., Liang, T., He, Z., Pang, J., Yu, D., Song, Hardt, M., and Levine, S. (eds.), NeurIPS, 2023. [L., Liu, Q., Zhou, M., Zhang, Z., Wang, R., Tu, Z., Mi, URL](https://md.research.elamur.ai/free/en/papers/chen-2024-think-that-much-2-3/) [http://dblp.uni-trier.de/db/conf/](http://dblp.uni-trier.de/db/conf/) H., and Yu, D. Do not think that much for 2+3=? on n… [↗](http://dblp.uni-trier.de/db/conf/)
11. Neubig, G. Don’t sleep on single-agent systems. [https://www.all-hands.dev/blog/](https://www.all-hands.dev/blog/) dont-sleep-on-single-agent-systems, 2024. Accessed: 2024-11-21. [↗](https://www.all-hands.dev/blog/)
12. Qwen. Qwq: Reflect deeply on the boundaries of the unknown, November 2024b. URL [https://qwenlm.github.io/blog/qwq-32b-preview/](https://qwenlm.github.io/blog/qwq-32b-preview/). [↗](https://qwenlm.github.io/blog/qwq-32b-preview/)
13. Rein, D., Hou, B. L., Stickland, A. C., Petty, J., Pang, R. Y., Dirani, J., Michael, J., and Bowman, S. R. [Gpqa: A graduate-level google-proof q&a benchmark](https://md.research.elamur.ai/free/en/papers/rein-2023-gpqa-graduate-level-google-proof/), 2023. URL https://arxiv.org/abs/2311.12022. [↗](https://arxiv.org/abs/2311.12022)
14. NovaSky. Sky-t1: Train your own o1 preview model within $450. [https://novasky-ai.github.io/posts/sky-t1](https://novasky-ai.github.io/posts/sky-t1), 2025. Accessed: 2025-01-09. [↗](https://novasky-ai.github.io/posts/sky-t1)
15. OpenAI. Openai function calling guide. [https://platform.openai.com/docs/guides/](https://platform.openai.com/docs/guides/) function-calling, 2024a. Accessed: 2024-11-21. [↗](https://platform.openai.com/docs/guides/)
16. Research, I. Swe agents: Empowering software development with ai agents. [https://research.ibm.com/blog/ibm-swe-agents](https://research.ibm.com/blog/ibm-swe-agents), 2024. Accessed: 2024-11-21. [↗](https://research.ibm.com/blog/ibm-swe-agents)
17. OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. [https://openai.com/index/](https://openai.com/index/) Russell, S. J. and Norvig, P. Artificial Intelligence: A Mod-gpt-4o-mini-advancing-cost-efficient-intelligence/, ern Approach. Prentice Hall, 1 edition, 1995. [↗](https://openai.com/index/)
18. ISBN 2024b. [Accessed: 2025-01-22](https://md.research.elamur.ai/free/en/papers/openai-2024-openai-o1-system-card/). 978-0-13-103805-9. Google-Books-ID: CUVeMwAA-CAAJ. OpenAI. Learning to reason with llms. [https://openai.com/index/](https://openai.com/index/) Smeyatsky, A. Agentic ai whitepaper, 2024. URL 2024c. learning-to-reason-with-llms/, https://www.linkedin.com/pulse/ [Online]. agentic-ai-whitepaper-al… [↗](https://openai.com/index/)
19. OpenAI. Gpt-4o mini model documentation. [https://platform.openai.com/docs/](https://platform.openai.com/docs/) models\#gpt-4o-mini, 2025b. Accessed: 2025-01-24. [↗](https://platform.openai.com/docs/)
20. Ma, R., Zheng, M., Qian, B., Shao, Y., Muennighoff, N., Zhang, Y., Hui, B., Lin, J., Brennan, R., Peng, H., Ji, H., and Neubig, G. [OpenHands: An Open Platform for AI Software Developers as Generalist Agents](https://md.research.elamur.ai/free/en/papers/wang-2024-openhands-open-platform-ai/), 2024b. URL https://arxiv.org/abs/2407.16741. [↗](https://arxiv.org/abs/2407.16741)
21. Wang, X., Li, B., Song, Y., Xu, F. F., Tang, X., Zhuge, M., Pan, J., Song, Y., Li, B., Singh, J., et al. [Openhands: An open platform for ai software developers as generalist agents](https://md.research.elamur.ai/free/en/papers/wang-2024-openhands-open-platform-ai/), 2024c.
22. OpenAI. Openai o1. [https://openai.com/o1/](https://openai.com/o1/), 2025c. Accessed: 2025-01-24. [↗](https://openai.com/o1/)
23. Phan, H. N., Nguyen, T. N., Nguyen, P. X., and Bui, N. D. Q. Hyperagent: Generalist software engineering agents to solve coding tasks at scale, 2024. URL https://arxiv.org/abs/2409.16299. — [arXiv:2409.16299](https://arxiv.org/abs/2409.16299) [↗](https://arxiv.org/abs/2409.16299)
24. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., and Zhou, D. [Chain-of-thought prompting elicits reasoning in large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-chain-thought-prompting-elicits-reasoning/), 2023. URL https://arxiv.org/abs/ 2201.11903. [↗](https://arxiv.org/abs/)
25. Qwen. Qwen2.5: A party of foundation models, September 2024a. URL [https://qwenlm.github.io/](https://qwenlm.github.io/) blog/qwen2.5/. [↗](https://qwenlm.github.io/)
26. Xia, C. S., Deng, Y., Dunn, S., and Zhang, L. [Agentless: Demystifying llm-based software engineering agents](https://md.research.elamur.ai/free/en/papers/xia-2024-agentless-demystifying-llm-based/), 2024. URL https://arxiv.org/abs/2407.01489. [↗](https://arxiv.org/abs/2407.01489)
27. Xu, F., Hao, Q., Zong, Z., Wang, J., Zhang, Y., Wang, J., Lan, X., Gong, J., Ouyang, T., Meng, F., Shao, C., Yan, Y., Yang, Q., Song, Y., Ren, S., Hu, X., Li, Y., Feng, J., Gao, C., and Li, Y. Towards large reasoning models: A survey on scaling llm reasoning capabilities, 2025. URL — [Towards Large Reasoning Models: A Survey of Reinforced Reasoning with Large Language Models](https://md.research.elamur.ai/free/en/papers/xu-2025-towards-large-reasoning-models-survey/) [https://arxiv.org…](https://arxiv.org…) [↗](https://arxiv.org…)
28. Yan, F., Mao, H., Ji, C. C.-J., Zhang, T., Patil, S. G., Stoica, I., and Gonzalez, J. E. Berkeley function calling leaderboard. 2024. [find](https://scholar.google.com/scholar?q=Yan%2C+F%2E%2C+Mao%2C+H%2E%2C+Ji%2C+C%2E+C%2E%2DJ%2E%2C+Zhang%2C+T%2E%2C+Patil%2C+S%2E+G%2E%2C+Stoica%2C+I%2E%2C+and+Gonzalez%2C+J%2E+E%2E+Berkeley+function+calling+leaderboard%2E+2024%2E)
29. Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024a. [find](https://scholar.google.com/scholar?q=Yang%2C+A%2E%2C+Yang%2C+B%2E%2C+Hui%2C+B%2E%2C+Zheng%2C+B%2E%2C+Yu%2C+B%2E%2C+Zhou%2C+C%2E%2C+Li%2C+C%2E%2C+Li%2C+C%2E%2C+Liu%2C+D%2E%2C+Huang%2C+F%2E%2C+et+al%2E+Qwen2+technical+report%2E+arXiv+preprint+arXiv%3A2407%2E10671%2C+2024a%2E)
30. Yang, J., Jimenez, C. E., Wettig, A., Lieret, K., Yao, S., Narasimhan, K., and Press, O. [Swe-agent: Agent-computer interfaces enable automated software engineering](https://md.research.elamur.ai/free/en/papers/yang-2024-swe-agent-agent-computer-interfaces/), 2024b. URL https://arxiv.org/abs/ 2405.15793. [↗](https://arxiv.org/abs/)
31. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T. L., Cao, Y., and Narasimhan, K. Tree of thoughts: Deliberate problem solving with large language models, 2023. URL https://arxiv.org/abs/2305.10601. [↗](https://arxiv.org/abs/2305.10601)
32. Zhang, W., Liao, J., Li, N., and Du, K. Agentic information retrieval, 2024. URL https://arxiv.org/abs/ 2410.09713. [↗](https://arxiv.org/abs/)
33. Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., Zhang, H., Gonzalez, J. E., and Stoica, I. [Judging llm-as-a-judge with mt-bench and chatbot arena](https://md.research.elamur.ai/free/en/papers/zheng-2023-judging-llm-judge-mt-bench/), 2023. URL https://arxiv.org/abs/2306.05685. [↗](https://arxiv.org/abs/2306.05685)
34. Zhong, T., Liu, Z., Pan, Y., Zhang, Y., Zhou, Y., Liang, S., Wu, Z., Lyu, Y., Shu, P., Yu, X., et al. [Evaluation of openai o1: Opportunities and challenges of agi](https://md.research.elamur.ai/free/en/papers/zhong-2024-evaluation-openai-o1-opportunities-challenges/). arXiv preprint arXiv:2409.18486, 2024.

*Updated 3 November 2026*